Multi-Valued Neuron with Sigmoid Activation Function Shin-Fu Wu 2013/6/21
MVN-sig Review
Outlines Stopping Criteria Parameter C Learning Rule for Parameter C Simulation Results Binary Classification MVN-P approach Simulation Results Complex-output Model Model Architecture Simulation Results Future Work
Stopping Criteria
Training acc. - EpochSquared error - Epoch
Training acc. - EpochSquared error - Epoch
Training acc. - EpochSquared error - Epoch
Parameter C
Simulation Results Wine Dataset MVN EpochSec.Acc MVN-sig (C=5) EpochSec.Acc MVN-sig (learned C) EpochSec.Acc
Simulation Results Glass Identification Dataset MVN EpochSec.Acc MVN-sig (C=5) EpochSec.Acc MVN-sig (learned C) EpochSec.Acc
Binary Classification MVN-P approach k=2, l=2, m=k*l=4 About 10% worse than MVN-P … WHY?
Simulation Results MVN-PMVN-sig-P Breast Cancer96.14%89% ~ 95.94% Parkinson's89.19%68.51% ~ 82.35% heart76.78%59.52% ~ 73.04%
Complex-output Model
Simulation Results Wine Dataset MVN EpochSec.Acc MVN-sig (C=5) EpochSec.Acc Complex MVN-sig (C=5) EpochSec.Acc
Simulation Results Iris Dataset MVN (96% trained) EpochSec.Acc MVN-sig (C=5) EpochSec.Acc Complex MVN-sig (C=5) EpochSec.Acc
Future Works Synthetic Data Analysis Why the binary classification failed? Why this model is feasible? Regression Problem How to solve regression problems? Multilayer Structure Construct MLMVN using complex-output MVN-sig How to choose the activation functions in the hidden layer?